GPT-6 Solves HellGates Custom CPU Challenge Previously Unsolvable by Humans and LLMs

HellGates Challenge Overview

The HellGates crackme, released in 2025, presented a challenge based on a custom 32-bit CPU designed in VHDL. This CPU was synthesized into a gate-level netlist incorporating multiple layers of obfuscation, anti-tamper measures, timing checks, and anti-debug features. The challenge required participants to reverse-engineer this complex system.

Previous Attempts and Failures

For over a year, neither human experts nor other prominent large language models, including Claude, ChatGPT, and DeepSeek, were able to solve the HellGates challenge. LLMs, even with hints, eventually declared the task "computationally infeasible with available resources," indicating the difficulty of the problem.

GPT-6's Breakthrough Solution

In September 2026, GPT-6 successfully solved the HellGates challenge in approximately 20-30 minutes on SRE-Bench. The solution exploited a weak cryptographic protection of the CPU's state, utilizing a side-channel/differential analysis attack. This allowed GPT-6 to replay runs, remove obfuscation layers, decrypt and re-encrypt registers, and ultimately dump the entire 1GB decrypted memory.

Implications for AI and Security

The successful resolution by GPT-6 highlights a significant advancement in AI's ability to analyze and overcome sophisticated hardware and software security challenges. This event demonstrates the evolving capabilities of advanced AI models in reverse engineering and bypassing complex protection mechanisms, even those designed to be difficult for both humans and prior AI systems.